Home

Aspect Based Sentiment Analysis Task

• Slot1
  • Given: sentence
  • Predict: E#A labels sentence
  • ex: "The food is expensive but service is good." => FOOD#PRICE, SERVICE#QUALITY
• Slot2
  • Given: sentence and E#A label
  • Predict: opinion target
  • ex: "The food is delicious", where is the FOOD#QUALITY? => "food"
• Slot3
  • Given: sentence, E#A label
  • Predict: polarity
  • ex: "The food is delicious", what is the polarity for "food" / FOOD#QUALITY? => "positive"

Baselines

• Slot1
  • task: Multi-label
  • feature: 1,000 most frequent unigrams of the training data excluding stop-words
  • model: SVM (linear kernel)
    • the label that over threshold prob is assigned (threshold = 0.2).
  • evaluation: f1 (micro-averaging)
• Slot2
  • task: Target Detection
  • model: dictionary ({"category": "word"}), it is made from opinions of training data. Search category by using dictionary, and the first match one is used. If the target does not exist, return the NULL.
  • evaluation: f1 (micro-averaging), discards NULL targets.
• Slot3:
  • task: Sentiment (0/1 prediction)
  • feature: 1,000 most frequent unigrams of the training data excluding stop-words + index of E&A cateogry
  • accuracy (number of correctly predicted polarity of the (gold) aspect categories / total number of the gold aspect categories)

Baseline scores

	SemEval-2016: Restaurant(EN)	chABSA-dataset
Slot1	59.928	44.772
Slot2	44.071	15.256
Slot3	76.484	75.886

Best model Score

	SemEval-2016: Restaurant(EN)	chABSA-dataset
Slot1	73.031	?
Slot2	72.34	?
Slot3	88.126	?

References

• Aspect based sentiment analysis
• SemEval-2014 Task 4: Aspect Based Sentiment Analysis
• SemEval-2015 Task 12: Aspect Based Sentiment Analysis
• SemEval-2016 Task 5: Aspect Based Sentiment Analysis
• Overview of Multilingual Opinion Analysis Task at NTCIR-7